MLIR — A Global Optimization and Dataflow Analysis

Table of Contents

In this article we’ll implement a global optimization pass, and show how to use the dataflow analysis framework to verify the results of our optimization.

The code for this article is in this pull request, and as usual the commits are organized to be read in order.

The noisy arithmetic problem

This demonstration is based on a simplified model of computation relevant to the HEIR project. You don’t need to be familiar with that project to follow this article, but if you’re wondering why someone would ever want the kind of optimization I’m going to write, that project is why.

The basic model is “noisy integer arithmetic.” That is, a program can have integer types of bounded width, and each integer is associated with some unknown (but bounded with high probability) symmetric random noise. You can imagine the “real” integer being the top 5 bits of a 32-bit integer, and the bottom 27 bits storing the noise. When a new integer is created, it magically has a random signed 12-bit integer added to it. When you apply operations to combine integers, the noise grows. Adding two integers adds their noise, and at worst you get one more bit of noise. Scaling an integer by a statically-known constant scales the noise by a constant. Multiplying two integers multiplies their noise values, and you get twice the bits of noise. As long as your program stays below 27 bits of noise, you can still “recover” the original 5-bit integer at the end of the program. Such a program is called legal, and otherwise, the output is random junk and the program is called illegal.

Finally, there is an expensive operation called reduce_noise that can explicitly reduce the noise of a noisy integer back to the base level of 12 bits. This operation has a statically known cost relative to the standard operations.

Note that starting from two noisy integers, each with 12 bits of noise, a single multiplication op brings you jarringly close to ruin. You would have at most 24 bits of noise, which is close to the maximum of 26. But the input IR we start with may have arbitrary computations that do not respect the noise limits. The goal of our optimization is to rewrite an MLIR region so that the noisy integer math never exceeds the noise limit at any given step.

A trivial way to do that would be to insert reduce_noise ops greedily, whenever an op would bring the noise of a value too high. Inserting such ops may be necessary, but a more suitable goal would be to minimize the overall cost of the program, subject to the constraint that the noisy arithmetic is legal. One could do this by inserting reduce_noise ops more strategically, or by rewriting the program to reduce the need for reduce_noise ops, or both. We’ll focus on the former: finding the best place to insert reduce_noise ops without rewriting the rest of the program.

The noisy dialect

We previously wrote about defining a new dialect, and the noisy dialect we created for this article has little new to show. This commit defines the types and ops, hard coding the 32-bit width and 5-bit semantic input type for simplicity, as well as the values of 12 bits of initial noise and 26 bits of max noise.

Note that the noise bound is not expressed on the type as an attribute. If it were, we’d run into a few problems: first, whenever you insert a reduce_noise op, you’d have to update the types on all the downstream ops. Second, it would prevent you from expressing control flow, since the noise bound cannot be statically inferred from the source code when there are two possible paths that could result in different noise values.

So instead, we need a way to compute the noise values, and associate them with each SSA value, and account for control flow. This is what an analysis pass is designed to do.

An analysis pass is just a class

The typical use of an analysis pass is to construct a data structure that encodes global information about a program, which can then be re-used during different parts of a pass. I imagined there would be more infrastructure around analysis passes in MLIR, but it’s quite simple. You define a C++ class with a constructor that takes an Operation *, and construct it basically whenever you need it. The only infrastructure for it involves storing and caching the constructed analysis within a pass, and understanding when an analysis needs to be recomputed (always between passes, by default).

By way of example, over at the HEIR project I made a simple analysis that chooses a unique variable name for every SSA value in a program, which I then used to generate code in an output language that needed variable names.

For this article we’ll see two analysis passes. One will formulate and solve the optimization problem that decides where to insert reduce_noise operations. This will be one of the “class that does anything” kind of analysis pass. The other analysis pass will rely on MLIR’s data flow analysis framework to propagate the noise model through the IR. This one will actually not require us to write an analysis from scratch, but instead will be implemented by means of the existing IntegerRangeAnalysis, which only requires us to implement an interface on each op that describes how the op affects the noise. This will be used in our pass to verify that the inserted reduce_noise operations ensure, if nothing else, that the noise never exceeds the maximum allowable noise.

We’ll start with the data flow analysis.

Reusing IntegerRangeAnalysis

Data flow analysis is a classical static analysis technique for propagating information through a program’s IR. It is one part of Frances Allen’s Turing Award. This article gives a good introduction and additional details, but I will paraphrase it briefly here in the context of IntegerRangeAnalysis.

The basic idea is that you want to get information about what possible values an integer-typed value can have at any point in a given program. If you see x = 7, then you know exactly what x is. If you see something like

func (%x : i8) { 
  %1 = arith.extsi %x : i8 to i32 
  %2 = arith.addi %x, %x : i32

then you know that %2 can be at most a signed 9-bit integer, because it started as an 8-bit integer, and adding two such integers together can’t fill up more than one extra bit.

In such cases, one can find optimizations, like the int-range-optimizations pass in MLIR, which looks at comparison ops arith.cmpi and determines if it can replace them with constants. It does this by looking at the integer range analysis for the two operands. E.g., given the op x > y, if you know y‘s maximum value is less than x‘s minimum value, then you can replace it with a constant true.

Computing the data flow analysis requires two ingredients called a transfer function and a join operation. The transfer function describes what the output integer range should be for a given op and a given set of input integer ranges. This can be an arbitrary function. The join operation describes how to combine two or more integer ranges when you get to a point at the program in which different branches of control flow merge. For example,

def fn(branch):
  x = 7
  if branch:
     y = x * x
     y = 2*x
  return y

The value of y just before returning cannot be known exactly, but in one branch you know it’s 14, and in another it’s 49. So the final value of y could be estimated as being in the range [14, 49]. Here the join function computes the smallest integer range containing both estimates. [Aside: it could instead use the set {14, 49} to be more precise, but that is not what IntegerRangeAnalysis happens to do]

In order for a data flow analysis to work properly, the values being propagated and the join function must together form a semilattice, which is a partially-ordered set in which every two elements have a least upper bound, that upper bound is computed by join, and join itself must be associative, commutative, and idempotent. For dataflow analysis, the semilattice must also be finite. This is often expressed by having distinct “top” and “bottom” elements as defaults. “Top” represents “could be anything,” and sometimes expresses that a more precise bound would be too computationally expensive to continue to track. “Bottom” usually represents an uninitialized value.

Once you have this, then MLIR provides a general algorithm to propagate values through the IR via a technique called Kildall’s method, which iteratively updates the SSA values, applying the transfer function and joining at the merging of control flow paths, until the process reaches a fixed point.

Here are MLIR’s official docs on dataflow analysis, and here is the RFC where the current data flow solver framework was introduced. In our situation, we want to use the solver framework with the existing IntegerRangeAnalysis, which only asks that we implement the transfer function by implementing InferIntRangeInterface on our ops.

This commit does just that. This requires adding the DeclareOpInterfaceMethods<InferIntRangeInterface> to all relevant ops. That in turn generates function declarations for

void MyOp::inferResultRanges(
    ArrayRef<ConstantIntRanges> inputRanges, SetIntRangeFn setResultRange);

The ConstantIntRange is a dataclass holding a min and max integer value. inputRanges represents the known bounds on the inputs to the operation in question, and SetIntRangeFn is the callback used to produce the result.

For example, for AddOp we can implement it as

ConstantIntRanges unionPlusOne(ArrayRef<ConstantIntRanges> inputRanges) {
  auto lhsRange = inputRanges[0];
  auto rhsRange = inputRanges[1];
  auto joined = lhsRange.rangeUnion(rhsRange);
  return ConstantIntRanges::fromUnsigned(joined.umin(), joined.umax() + 1);

void AddOp::inferResultRanges(ArrayRef<ConstantIntRanges> inputRanges,
                              SetIntRangeFn setResultRange) {
  setResultRange(getResult(), unionPlusOne(inputRanges));

A MulOp is similarly implemented by summing the maxes. Meanwhile, EncodeOp and ReduceNoiseOp each set the initial range to [0, 12]. So the min will always be zero, and we really only care about the max.

The next commit defines an empty pass that will contain our analyses and optimizations, and this commit shows how the integer range analysis is used to validate an IR’s noise growth. In short, you load the IntegerRangeAnalysis and its dependent DeadCodeAnalysis, run the solver, and then walk the IR, asking the solver via lookupState to give the resulting value range for each op’s result, and comparing it against the maximum.

void runOnOperation() {
    Operation *module = getOperation();

    DataFlowSolver solver;
    if (failed(solver.initializeAndRun(module)))

    auto result = module->walk([&](Operation *op) {
      if (!llvm::isa<noisy::AddOp, noisy::SubOp, noisy::MulOp,
                     noisy::ReduceNoiseOp>(*op)) {
        return WalkResult::advance();
      const dataflow::IntegerValueRangeLattice *opRange =
      if (!opRange || opRange->getValue().isUninitialized()) {
            << "Found op without a set integer range; did the analysis fail?";
        return WalkResult::interrupt();

      ConstantIntRanges range = opRange->getValue().getValue();
      if (range.umax().getZExtValue() > MAX_NOISE) {
        op->emitOpError() << "Found op after which the noise exceeds the "
                             "allowable maximum of "
                          << MAX_NOISE
                          << "; it was: " << range.umax().getZExtValue()
                          << "\n";
        return WalkResult::interrupt();

      return WalkResult::advance();

    if (result.wasInterrupted())

Finally, in this commit we add a test that exercises it:

func.func @test_op_syntax() -> i5 {
  %0 = arith.constant 3 : i5
  %1 = arith.constant 4 : i5
  %2 = noisy.encode %0 : i5 -> !noisy.i32
  %3 = noisy.encode %1 : i5 -> !noisy.i32
  %4 = noisy.mul %2, %3 : !noisy.i32
  %5 = noisy.mul %4, %4 : !noisy.i32
  %6 = noisy.mul %5, %5 : !noisy.i32
  %7 = noisy.mul %6, %6 : !noisy.i32
  %8 = noisy.decode %7 : !noisy.i32 -> i5
  return %8 : i5

Running tutorial-opt --noisy-reduce-noise on this file produces the following error:

 error: 'noisy.mul' op Found op after which the noise exceeds the allowable maximum of 26; it was: 48

  %5 = noisy.mul %4, %4 : !noisy.i32
mlir-tutorial/tests/noisy_reduce_noise.mlir:11:8: note: see current operation: %5 = "noisy.mul"(%4, %4) : (!noisy.i32, !noisy.i32) -> !noisy.i32

And if you run in debug mode with --debug --debug-only=int-range-analysis, you will see the per-op propagations printed to the terminal

$  bazel run tools:tutorial-opt -- --noisy-reduce-noise-optimizer $PWD/tests/noisy_reduce_noise.mlir --debug --debug-only=int-range-analysis
Inferring ranges for %c3_i5 = arith.constant 3 : i5
Inferred range unsigned : [3, 3] signed : [3, 3]
Inferring ranges for %c4_i5 = arith.constant 4 : i5
Inferred range unsigned : [4, 4] signed : [4, 4]
Inferring ranges for %0 = noisy.encode %c3_i5 : i5 -> !noisy.i32
Inferred range unsigned : [0, 12] signed : [0, 12]
Inferring ranges for %1 = noisy.encode %c4_i5 : i5 -> !noisy.i32
Inferred range unsigned : [0, 12] signed : [0, 12]
Inferring ranges for %2 = noisy.mul %0, %1 : !noisy.i32
Inferred range unsigned : [0, 24] signed : [0, 24]
Inferring ranges for %3 = noisy.mul %2, %2 : !noisy.i32
Inferred range unsigned : [0, 48] signed : [0, 48]
Inferring ranges for %4 = noisy.mul %3, %3 : !noisy.i32
Inferred range unsigned : [0, 96] signed : [0, 96]
Inferring ranges for %5 = noisy.mul %4, %4 : !noisy.i32
Inferred range unsigned : [0, 192] signed : [0, 192]

As a quick aside, there was one minor upstream problem preventing me from reusing IntegerRangeAnalysis, which I patched in This means I also had to update the LLVM commit hash used by this project in this commit.

An ILP optimization pass

Next, we build an analysis that solves a global optimization to insert reduce_noise ops efficiently. As mentioned earlier, this is a “do anything” kind of analysis, so we put all of the logic into the analysis’s construtor.

[Aside: I wouldn’t normally do this, because constructors don’t have return values so it’s hard to signal failure; but the API for the analysis specifies the constructor takes as input the Operation * to analyze, and I would expect any properly constructed object to be “ready to use.” Maybe someone who knows C++ better will comment and shed some wisdom for me.]

This commit sets up the analysis shell and interface.

class ReduceNoiseAnalysis {
  ReduceNoiseAnalysis(Operation *op);
  ~ReduceNoiseAnalysis() = default;

  /// Return true if a reduce_noise op should be inserted after the given
  /// operation, according to the solution to the optimization problem.
  bool shouldInsertReduceNoise(Operation *op) const {
    return solution.lookup(op);

  llvm::DenseMap<Operation *, bool> solution;

This commit adds a workspace dependency on Google’s or-tools package (“OR” stands for Operations Research here, a.k.a. discrete optimization), which comes bundled with a number of nice solvers, and an API for formulating optimization problems. And this commit implements the actual solver model.

Now this model is quite a bit of code, and this article is not the best place to give a full-fledged introduction to linear programming, modeling techniques, or the OR-tools API. What I’ll do instead is explain the model in detail here, give a few small notes on how that translates to the OR-tools C++ API. If you want a gentler background on linear programming, see my article series about diet optimization (part 1, part 2).

All linear programs specify a linear function as an objective to minimize, along with a set of linear equalities and inequalities that constrain the solution. In a standard linear program, the variables must be continuously valued. In a mixed-integer linear program, some of those variables are allowed to be discrete integers, which, it turns out, makes it possible to solve many more problems, but requires completely different optimization techniques and may result in exponentially slow runtime. So many techniques in operations research relate to modeling a problem in such a way that the number of integer variables is relatively small.

Our linear model starts by defining some basic variables. Some variables in the model represent “decisions” that we can make, and others represent “state” that reacts to the decisions via constraints.

  • For each operation $x$, a $\{0, 1\}$-valued variable $\textup{InsertReduceNoise}_x$. Such a variable is 1 if and only if we insert a reduce_noise op after the operation $x$.
  • For each SSA-value $v$ that is input or output to a noisy op, a continuous-valued variable $\textup{NoiseAt}_v$. This represents the upper bound of the noise at value $v$.

In particular, the solver’s performance will get worse as the number of binary variables increases, which in this case corresponds to the number of noisy ops.

The objective function, with a small caveat explained later, is simply the sum of the decision variables, and we’d like to minimize it. Each reduce_noise op is considered equally expensive, and there is no special nuance here about scheduling them in parallel or in serial.

Next, we add constraints. First, $0 \leq \textup{NoiseAt}_v \leq 26$, which asserts that no SSA value can exceed the max noise. Second, we need to enforce that an encode op fixes the noise of its output to 12, i.e., for each encode op $x$, we add the constraint $\textup{NoiseAt}_{\textup{result}(x)} = 12$.

Finally, we need constraints that say that if you choose to insert a reduce_noise op, then the noise is reset to 12, otherwise it is set to the appropriate function of the inputs. This is where the modeling gets a little tricky, but multiplication is easier so let’s start there.

Fix a multiplication op $x$, its two input SSA values $\textup{LHS}, \textup{RHS}$, and its output $\textup{RES}$. As a piecewise function, we want a constraint like:

\[ \textup{NoiseAt}_\textup{RES} = \begin{cases} \textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS} & \text{ if } \textup{InsertReduceNoise}_x = 0 \\ 12 & \text{ if } \textup{InsertReduceNoise}_x = 1 \end{cases} \]

This isn’t linear, but we can combine the two branches to

\[ \begin{aligned} \textup{NoiseAt}_\textup{RES} &= (1 – \textup{ InsertReduceNoise}_x) (\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS}) \\ & + 12 \textup{ InsertReduceNoise}_x \end{aligned} \]

This does the classic trick of using a bit as a controlled multiplexer, but it’s still not linear. We can make it linear, however, by replacing this one constraint with four constraints, and an auxiliary constant $C=100$ that we know is larger than the possible range of values that the $\textup{NoiseAt}_v$ variables can attain. Those four linear constraints are:

\[ \begin{aligned}
\textup{NoiseAt}_\textup{RES} &\geq 12 \textup{ InsertReduceNoise}_x \\
\textup{NoiseAt}_\textup{RES} &\leq 12 + C(1 – \textup{InsertReduceNoise}_x) \\
\textup{NoiseAt}_\textup{RES} &\geq (\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS}) – C \textup{ InsertReduceNoise}_x \\
\textup{NoiseAt}_\textup{RES} &\leq (\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS}) + C \textup{ InsertReduceNoise}_x \\
\end{aligned} \]

Setting the decision variable to zero results in the first two equations being trivially satisfied. Setting it to 1 causes the first two equations to be equivalent to $\textup{NoiseAt}_\textup{RES} = 12$. Likewise, the second two constraints are trivial when the decision variable is 1, and force the output noise to be equal to the sum of the two input noises when set to zero.

The addition op is handled similarly, except that the term $(\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS})$ is replaced by something non-linear, namely $1 + \max(\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS})$. We can still handle that, but it requires an extra modeling trick. We introduce a new variable $Z_x$ for each add op $x$, and two constraints:

\[ \begin{aligned} Z_v &\geq 1 + \textup{NoiseAt}_{LHS} \\ Z_v &\geq 1 + \textup{NoiseAt}_{RHS} \end{aligned} \]

Together these ensure that $Z_v$ is at least 1 plus the max of the two input noises, but it doesn’t force equality. To achieve that, we add $Z_v$ to the minimization objective (alongside the sum of the decision variables) with a small penalty to ensure the solver tries to minimize them. Since they have trivially minimal values equal to “1 plus the max,” the solver will have no trouble optimizing them, and this will be effectively an equality constraint.

[Aside: Whenever you do this trick, you have to convince yourself that the solver won’t somehow be able to increase $Z_v$ as a trade-off against lower values of other objective terms, and produce a lower overall objective value. Solvers are mischievous and cannot be trusted. In our case, there is no risk: if you were to increase $Z_v$ below its minimum value, that would only increase the noise propagation through add ops, meaning the solver would have to compensate by potentially adding even more reduce_noise ops!]

Then, the constraint for an add op uses $Z_v$ in place of $(\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS})$ the mul op.

The only other minor aspect of this solver model is that these constraints enforce the consistency of the noise propagation after a reduce_noise op may be added, but if a reduce_noise op is added, it doesn’t necessarily enforce the noise growth of the output of the op before it’s input to reduce_noise. We can achieve this by adding new constraints expressing $(\textup{NoiseAt}_{LHS} + \textup{NoiseAt}_{RHS}) leq 26$ and $Z_v \leq 26$ for multiplication and addition ops, respectively.

When converting this to the OR-tools C++ API, as we did in this commit, a few minor things to note:

  • You can specify upper and lower bounds on a variable at variable creation time, rather than as separate constraints. You’ll see this in solver->MakeNumVar(min, max, name).
  • Constraints must be specified in the form min <= expr <= max, where min and max are constants and expr is a linear combination of variables, meaning that one has to manually re-arrange and simplify all the equations above so the variables are all on one side and the constants on the other. (The OR-tools Python API is more expressive, but we don’t have it here.)
  • The constraints and the objective are specified by SetCoefficient, which sets the coefficient of a variable in a linear combination one at a time.

Finally, this commit implements the part of the pass that uses the solver’s output to insert new reduce_noise ops. And this commit adds some more tests.

An example of its use:

// This test checks that the solver can find a single insertion point
// for a reduce_noise op that handles two branches, each of which would
// also need a reduce_noise op if handled separately.
func.func @test_single_insertion_branching() -> i5 {
  %0 = arith.constant 3 : i5
  %1 = arith.constant 4 : i5
  %2 = noisy.encode %0 : i5 -> !noisy.i32
  %3 = noisy.encode %1 : i5 -> !noisy.i32
  // Noise: 12
  %4 = noisy.mul %2, %3 : !noisy.i32
  // Noise: 24

  // branch 1
  %b1 = noisy.add %4, %3 : !noisy.i32
  // Noise: 25
  %b2 = noisy.add %b1, %3 : !noisy.i32
  // Noise: 25
  %b3 = noisy.add %b2, %3 : !noisy.i32
  // Noise: 26
  %b4 = noisy.add %b3, %3 : !noisy.i32
  // Noise: 27

  // branch 2
  %c1 = noisy.sub %4, %2 : !noisy.i32
  // Noise: 25
  %c2 = noisy.sub %c1, %3 : !noisy.i32
  // Noise: 25
  %c3 = noisy.sub %c2, %3 : !noisy.i32
  // Noise: 26
  %c4 = noisy.sub %c3, %3 : !noisy.i32
  // Noise: 27

  %x1 = noisy.decode %b4 : !noisy.i32 -> i5
  %x2 = noisy.decode %c4 : !noisy.i32 -> i5
  %x3 = arith.addi %x1, %x2 : i5
  return %x3 : i5

And the output:

  func.func @test_single_insertion_branching() -> i5 {
    %c3_i5 = arith.constant 3 : i5
    %c4_i5 = arith.constant 4 : i5
    %0 = noisy.encode %c3_i5 : i5 -> !noisy.i32
    %1 = noisy.encode %c4_i5 : i5 -> !noisy.i32
    %2 = noisy.mul %0, %1 : !noisy.i32
    %3 = noisy.reduce_noise %2 : !noisy.i32
    %4 = noisy.add %3, %1 : !noisy.i32
    %5 = noisy.add %4, %1 : !noisy.i32
    %6 = noisy.add %5, %1 : !noisy.i32
    %7 = noisy.add %6, %1 : !noisy.i32
    %8 = noisy.sub %3, %0 : !noisy.i32
    %9 = noisy.sub %8, %1 : !noisy.i32
    %10 = noisy.sub %9, %1 : !noisy.i32
    %11 = noisy.sub %10, %1 : !noisy.i32
    %12 = noisy.decode %7 : !noisy.i32 -> i5
    %13 = noisy.decode %11 : !noisy.i32 -> i5
    %14 = arith.addi %12, %13 : i5
    return %14 : i5

Earthmover Distance

Problem: Compute distance between points with uncertain locations (given by samples, or differing observations, or clusters).

For example, if I have the following three “points” in the plane, as indicated by their colors, which is closer, blue to green, or blue to red?


It’s not obvious, and there are multiple factors at work: the red points have fewer samples, but we can be more certain about the position; the blue points are less certain, but the closest non-blue point to a blue point is green; and the green points are equally plausibly “close to red” and “close to blue.” The centers of masses of the three sample sets are close to an equilateral triangle. In our example the “points” don’t overlap, but of course they could. And in particular, there should probably be a nonzero distance between two points whose sample sets have the same center of mass, as below. The distance quantifies the uncertainty.


All this is to say that it’s not obvious how to define a distance measure that is consistent with perceptual ideas of what geometry and distance should be.

Solution (Earthmover distance): Treat each sample set $ A$ corresponding to a “point” as a discrete probability distribution, so that each sample $ x \in A$ has probability mass $ p_x = 1 / |A|$. The distance between $ A$ and $ B$ is the optional solution to the following linear program.

Each $ x \in A$ corresponds to a pile of dirt of height $ p_x$, and each $ y \in B$ corresponds to a hole of depth $ p_y$. The cost of moving a unit of dirt from $ x$ to $ y$ is the Euclidean distance $ d(x, y)$ between the points (or whatever hipster metric you want to use).

Let $ z_{x, y}$ be a real variable corresponding to an amount of dirt to move from $ x \in A$ to $ y \in B$, with cost $ d(x, y)$. Then the constraints are:

  • Each $ z_{x, y} \geq 0$, so dirt only moves from $ x$ to $ y$.
  • Every pile $ x \in A$ must vanish, i.e. for each fixed $ x \in A$, $ \sum_{y \in B} z_{x,y} = p_x$.
  • Likewise, every hole $ y \in B$ must be completely filled, i.e. $ \sum_{y \in B} z_{x,y} = p_y$.

The objective is to minimize the cost of doing this: $ \sum_{x, y \in A \times B} d(x, y) z_{x, y}$.

In python, using the ortools library (and leaving out a few docstrings and standard import statements, full code on Github):

from ortools.linear_solver import pywraplp

def earthmover_distance(p1, p2):
    dist1 = {x: count / len(p1) for (x, count) in Counter(p1).items()}
    dist2 = {x: count / len(p2) for (x, count) in Counter(p2).items()}
    solver = pywraplp.Solver('earthmover_distance', pywraplp.Solver.GLOP_LINEAR_PROGRAMMING)

    variables = dict()

    # for each pile in dist1, the constraint that says all the dirt must leave this pile
    dirt_leaving_constraints = defaultdict(lambda: 0)

    # for each hole in dist2, the constraint that says this hole must be filled
    dirt_filling_constraints = defaultdict(lambda: 0)

    # the objective
    objective = solver.Objective()

    for (x, dirt_at_x) in dist1.items():
        for (y, capacity_of_y) in dist2.items():
            amount_to_move_x_y = solver.NumVar(0, solver.infinity(), 'z_{%s, %s}' % (x, y))
            variables[(x, y)] = amount_to_move_x_y
            dirt_leaving_constraints[x] += amount_to_move_x_y
            dirt_filling_constraints[y] += amount_to_move_x_y
            objective.SetCoefficient(amount_to_move_x_y, euclidean_distance(x, y))

    for x, linear_combination in dirt_leaving_constraints.items():
        solver.Add(linear_combination == dist1[x])

    for y, linear_combination in dirt_filling_constraints.items():
        solver.Add(linear_combination == dist2[y])

    status = solver.Solve()
    if status not in [solver.OPTIMAL, solver.FEASIBLE]:
        raise Exception('Unable to find feasible solution')

    return objective.Value()

Discussion: I’ve heard about this metric many times as a way to compare probability distributions. For example, it shows up in an influential paper about fairness in machine learning, and a few other CS theory papers related to distribution testing.

One might ask: why not use other measures of dissimilarity for probability distributions (Chi-squared statistic, Kullback-Leibler divergence, etc.)? One answer is that these other measures only give useful information for pairs of distributions with the same support. An example from a talk of Justin Solomon succinctly clarifies what Earthmover distance achieves

Screen Shot 2018-03-03 at 6.11.00 PM.png

Also, why not just model the samples using, say, a normal distribution, and then compute the distance based on the parameters of the distributions? That is possible, and in fact makes for a potentially more efficient technique, but you lose some information by doing this. Ignoring that your data might not be approximately normal (it might have some curvature), with Earthmover distance, you get point-by-point details about how each data point affects the outcome.

This kind of attention to detail can be very important in certain situations. One that I’ve been paying close attention to recently is the problem of studying gerrymandering from a mathematical perspective. Justin Solomon of MIT is a champion of the Earthmover distance (see his fascinating talk here for more, with slides) which is just one topic in a field called “optimal transport.”

This has the potential to be useful in redistricting because of the nature of the redistricting problem. As I wrote previously, discussions of redistricting are chock-full of geometry—or at least geometric-sounding language—and people are very concerned with the apparent “compactness” of a districting plan. But the underlying data used to perform redistricting isn’t very accurate. The people who build the maps don’t have precise data on voting habits, or even locations where people live. Census tracts might not be perfectly aligned, and data can just plain have errors and uncertainty in other respects. So the data that district-map-drawers care about is uncertain much like our point clouds. With a theory of geometry that accounts for uncertainty (and the Earthmover distance is the “distance” part of that), one can come up with more robust, better tools for redistricting.

Solomon’s website has a ton of resources about this, under the names of “optimal transport” and “Wasserstein metric,” and his work extends from computing distances to computing important geometric values like the barycenter, computational advantages like parallelism.

Others in the field have come up with transparency techniques to make it clearer how the Earthmover distance relates to the geometry of the underlying space. This one is particularly fun because the explanations result in a path traveled from the start to the finish, and by setting up the underlying metric in just such a way, you can watch the distribution navigate a maze to get to its target. I like to imagine tiny ants carrying all that dirt.

Screen Shot 2018-03-03 at 6.15.50 PM.png

Finally, work of Shirdhonkar and Jacobs provide approximation algorithms that allow linear-time computation, instead of the worst-case-cubic runtime of a linear solver.

Boolean Logic in Polynomials

Problem: Express a boolean logic formula using polynomials. I.e., if an input variable $ x$ is set to $ 0$, that is interpreted as false, while $ x=1$ is interpreted as true. The output of the polynomial should be 0 or 1 according to whether the formula is true or false as a whole.

Solution: You can do this using a single polynomial.

Illustrating with an example: the formula is $ \neg[(a \vee b) \wedge (\neg c \vee d)]$ also known as

not((a or b) and (not c or d))

The trick is to use multiplication for “and” and $ 1-x$ for “not.” So $ a \wedge b$ would be $ x_1 x_2$, and $ \neg z$ would be $ 1-z$. Indeed, if you have two binary variables $ x$ and $ y$ then $ xy$ is 1 precisely when both are 1, and zero when either variable is zero. Likewise, $ 1-x = 1$ if $ x$ is zero and zero if $ x$ is one.

Combine this with deMorgan’s rule to get any formula. $ a \vee b = \neg(\neg a \wedge \neg b)$ translates to $ 1 – (1-a)(1-b)$. For our example above,

$ \displaystyle f(x_1, x_2, x_3, x_4) = 1 – (1 – (1-a)(1-b))(1 – c(1-d))$

Which expands to

$ \displaystyle 1 – a – b + ab + (1-d)(ac + bc – abc)$

If you plug in $ a = 1, b = 0, c = 1, d = 0$ you get True in the original formula (because “not c or d” is False), and likewise the polynomial is

$ \displaystyle 1 – 1 – 0 + 0 + (1-0)(1 + 0 – 0) = 1$

You can verify the rest work yourself, using the following table as a guide:

0, 0, 0, 0 -&gt; 1
0, 0, 0, 1 -&gt; 1
0, 0, 1, 0 -&gt; 1
0, 0, 1, 1 -&gt; 1
0, 1, 0, 0 -&gt; 0
0, 1, 0, 1 -&gt; 0
0, 1, 1, 0 -&gt; 1
0, 1, 1, 1 -&gt; 0
1, 0, 0, 0 -&gt; 0
1, 0, 0, 1 -&gt; 0
1, 0, 1, 0 -&gt; 1
1, 0, 1, 1 -&gt; 0
1, 1, 0, 0 -&gt; 0
1, 1, 0, 1 -&gt; 0
1, 1, 1, 0 -&gt; 1
1, 1, 1, 1 -&gt; 0

Discussion: This trick is used all over CS theory to embed boolean logic within polynomials, and it makes the name “boolean algebra” obvious, because it’s just a subset of normal algebra.

Moreover, since boolean satisfiability—the problem of algorithmically determining if a boolean formula has a satisfying assignment (a choice of variables evaluating to true)—is NP-hard, this can be used to show certain problems relating to multivariable polynomials is also hard. For example, finding roots of multivariable polynomials (even if you knew nothing about algebraic geometry) is hard because you’d run into NP-hardness by simply considering the subset of polynomials coming from boolean formulas.

Here’s a more interesting example, related to the kinds of optimization problems that show up in modern machine learning. Say you want to optimize a polynomial $ f(x)$ subject to a set of quadratic equality constraints. This is NP-hard. Here’s why.

Let $ \varphi$ be a boolean formula, and $ f_\varphi$ its corresponding polynomial. First, each variable $ x_i$ used in the polynomial can be restricted to binary values via the constraint $ x_i(x_i – 1) = 0$.

You can even show NP-hardness if the target function to optimize is only quadratic. As an exercise, one can express the subset sum problem as a quadratic programming problem using similar choices for the constraints. According to this writeup you even express subset sum as a quadratic program with linear constraints.

The moral of the story is simply that multivariable polynomials can encode arbitrary boolean logic.

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.


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.


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.


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.


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.


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 \|$.


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$.


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.



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!